betaFeedback

WorksEngineering

Distance-Certified Safe Control in Unknown Environments via Spatio-Temporal Composition of Gaussian Processes

Abstract

In this paper, we propose a distance-certified safety control framework for unknown and dynamic environments. The proposed strategy relies on a spatio-temporal composition of Gaussian Process (GP) barriers from environment point clouds. A point-cloud-dependent covariance kernel is constructed to smoothly combine individual GP representations, yielding a continuously differentiable posterior mean and a diagonal observation Gram matrix that avoids dense covariance inversions. We then derive a closed-form calibration of the GP length scale accounting for spatial approximation errors and finite point-cloud coverage, guaranteeing a prescribed minimum distance from the true obstacle boundary. To handle moving obstacles and updated observations, the scan-based barriers are certified over finite validity horizons and smoothly combined into a time-varying Control Barrier Function (CBF). Under problem feasibility conditions, the associated CBF-based safety control guarantees forward invariance of the certified safe set, thereby ensuring the prescribed distance from the obstacles. Unicycle simulations with static and moving obstacles illustrate the effectiveness of the proposed approach.

Introduction

The increasing complexity of autonomous systems operating in real-world unstructured environments has led to extensive research in the field of safety-critical control, whose goal is to enforce safety specifications while preserving the performance of a nominal control policy (see, as a reference, the recent survey [1]). Among other methods, Control Barrier Functions (CBFs) provide a systematic framework for this purpose by establishing forward invariance of prescribed safe sets [2, 3]. These methods have found widespread applications in robotics, including collision avoidance in multi-robot systems [4], robotic manipulation and human-robot interaction [5, 6, 7], as well as autonomous vehicles [8]. Nevertheless, their implementation requires a suitable mathematical representation of the safe set. In obstacle avoidance problems, such a representation depends on the geometry of the surrounding environment, which is often unknown a priori.

To address this limitation, several approaches have investigated the construction of CBFs from available measurements and/or learned representations. These include learning barriers from expert demonstrations [9], accounting for perception errors [10], and constructing implicit obstacle representations from range measurements [11]. More directly, point-cloud-based methods have been proposed to construct local barrier functions [12], enable safe navigation in unknown environments [13], and efficiently compute distance functions for nonconvex obstacles [14].

Among other approaches, Gaussian Processes (GPs) provide a flexible framework for representing unknown functions from data, while also quantifying the associated uncertainty [15]. Their integration with CBFs has been successfully explored in several settings, including multi-robot collision avoidance [16], safe learning in unknown environments [17], and online LiDAR-based navigation [18]. Further developments have considered GP-based estimation of uncertain dynamics for high-order safety filtering [19], and dynamic GP-based barrier functions for moving obstacles [20]. Alternative implicit safety representations, such as the Poisson safety function [21], have also been proposed. Despite these advancements, establishing a direct relationship between the level sets of a GP-based obstacle representation and a user-defined minimum distance to the obstacles remains challenging, especially in the presence of finite boundary sampling.

An additional difficulty arises when the environment evolves over time. Moving obstacles and periodically updated measurements require safety certificates to account for both the evolution of the underlying constraints and the replacement of previously acquired information. Time-varying CBFs extend the standard forward-invariance formulation to evolving safe sets [22, 23, 24], with applications ranging from environment-adaptive assistance [25] to mobile-robot navigation [26] and dynamic obstacle avoidance [27]. In parallel, smooth compositions of multiple barrier functions have been investigated to accommodate multiple safety requirements [28]. Relevant to perception-driven control, a time-varying soft-maximum CBF construction has been proposed in [29]. However, the smooth temporal composition of barrier functions does not, by itself, establish a prescribed minimum distance from the true obstacle boundary. These considerations highlight two closely related challenges in perception-based safety-critical control: (i) learning, from finite boundary sampling, a smooth obstacle representation with distance-certified level sets, and (ii) providing a distance-aware safety policy in the presence of moving obstacles. Addressing both aspects requires connecting the geometry of obstacle point-cloud observations with time-varying safety control certificates.

In this paper, we propose a perception-aware safety controller based on Gaussian Processes with distance-certified guarantees. The resulting barrier function is guaranteed to maintain a prescribed minimum distance from the true obstacle boundary, despite

The work was partially funded by the FISA-2023-00210 project APACHE - CUP J53C25000520001, by Next Generation EU, National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 3.3; CUP J33C24001490009, and by IMA S.p.A. The corresponding author is Davide Zarabini.

The authors are with the Department of Electrical, Electronic and Information Engineering, University of Bologna, 40136, Bologna, Italy (e-mail: [email protected], [email protected], [email protected]). finite point-cloud sampling and obstacle motion. Specifically, the main contribution of the paper is threefold. First, we introduce a novel positive-semidefinite covariance kernel that combines individual observations through smooth spatial weights. The resulting GP admits an analytical, continuously differentiable posterior mean and a diagonal observation Gram matrix, which avoids dense conditioning inversions and enables efficient online updates. Second, we establish a deterministic relationship between the GP posterior superlevel set and the complete obstacle boundary. By deriving a bound on the spatial weighting approximation and accounting for finite point-cloud coverage, we obtain a closed-form calibration of the GP length scale that guarantees a user-defined minimum distance to the true obstacle boundary. We further extend this result to moving obstacles by accounting for bounded obstacle motion and finite scan-validity horizons. Third, we develop a time-varying safety controller that combines the certified scan-based barriers through smooth temporal weights, and we establish forward invariance of the safe set and a minimum distance to the true obstacle boundary.

The paper is organized as follows. Section II reviews some preliminaries on Gaussian Processes and Control Barrier Functions. Section III introduces the considered framework and formulates the problem. Section IV presents the main theoretical result of the paper and the corresponding algorithmic implementation, while Section V provides the supporting theoretical analysis, including the proof of the main result. Section VI evaluates the proposed approach with numerical simulations. Formal proofs of the supporting theoretical results are provided in the Appendix.

Notation: R≥0\mathbb{R}_{\ge0} and N≥0\mathbb{N}_{\ge0} denote the nonnegative real numbers and nonnegative integers, respectively. We denote by ∇f\nabla f the gradient of a differentiable function ff, and by ∂tf\partial_t f its partial derivative with respect to time tt. A continuous, strictly increasing function α:R→R\alpha: \mathbb{R} \to\mathbb{R} with α(0)=0\alpha(0) = 0 is of extended class K\mathcal{K}. The indicator 1i=j\mathbf{1}_{i=j} equals 1 if i=ji = j and 0 otherwise. The floor function ⌊r⌋\lfloor r\rfloor denotes the greatest integer less than or equal to rr. The Euclidean norm of a vector x∈Rnx \in\mathbb{R}^n is denoted by ∥x∥\lVert x\rVert. For a nonempty set AA, the distance from a point xx to the set AA is denoted by dist⁡(x,A):=inf⁡a∈A∥x−a∥\operatorname{dist}(x,A) := \inf_{a\in A} \lVert x-a\rVert. For nonempty subsets AA, BB of the same Euclidean space, their Hausdorff distance is defined as dH(A,B):=max⁡{sup⁡a∈Adist⁡(a,B),sup⁡b∈Bdist⁡(b,A)}d_H(A,B) := \max\{\sup_{a\in A}\operatorname{dist}(a,B),\sup_{b\in B}\operatorname{dist}(b,A)\}. For any τ>0\tau> 0, define the continuously differentiable, compactly supported window χτ(s):=sin⁡2(πsτ)1[0,τ](s)\chi_\tau(s) := \sin^2\left(\frac{\pi s}{\tau}\right)\mathbf{1}_{[0,\tau]}(s).

Preliminaries

Preliminaries on Gaussian Processes

A Gaussian Process (GP) is a stochastic process defining a distribution over functions [15]. We denote a GP by

f(x)∼GP(μ(x),k(x,x′)),(1)f(x) \sim\mathcal{GP}(\mu(x),k(x,x')), \tag*{(1)}

where μ:Rn→R\mu: \mathbb{R}^n \to\mathbb{R} is the mean function and k:Rn×Rn→Rk : \mathbb{R}^n \times\mathbb{R}^n \to\mathbb{R} is the covariance function. The covariance function must be symmetric and positive semidefinite, ensuring that the covariance matrix of any finite collection of function values is positive semidefinite. A critical property of GPs is that they are closed under conditioning, allowing us to compute the posterior distribution of the function given observed data. To this end, consider a dataset D={(xi,yi)}i=1N\mathcal{D} = \{(x_i,y_i)\}_{i=1}^{N} of NN input-output pairs, where yi=f(xi)+ϵiy_i = f(x_i) + \epsilon_i and ϵi∼N(0,σn2)\epsilon_i \sim\mathcal{N}(0,\sigma_n^2) is i.i.d. Gaussian noise. The posterior distribution of ff at a new input x∗x_* is given by

μ∗(x∗)=μ(x∗)+k(x∗,X)(K+σn2I)−1(y−μ(X)),(2)\mu_*(x_*) = \mu(x_*) + k(x_*,X)(K+\sigma_n^2 I)^{-1}(y-\mu(X)), \tag*{(2)}
σ∗2(x∗)=k(x∗,x∗)−k(x∗,X)(K+σn2I)−1k(X,x∗),(3)\sigma_*^2(x_*) = k(x_*,x_*) - k(x_*,X)(K+\sigma_n^2 I)^{-1}k(X,x_*), \tag*{(3)}

where μ(X)=[μ(x1),…,μ(xN)]T\mu(X) = [\mu(x_1),\ldots,\mu(x_N)]^\mathsf{T} is the vector of prior mean evaluations, X=[x1,…,xN]TX = [x_1,\ldots,x_N]^\mathsf{T} is the matrix of training inputs, y=[y1,…,yN]Ty = [y_1,\ldots,y_N]^\mathsf{T} is the vector of training outputs, and K∈RN×NK \in\mathbb{R}^{N\times N} is the Gram covariance matrix with entries Kij=k(xi,xj)K_{ij} = k(x_i,x_j). A common choice for the covariance function is the squared-exponential kernel, defined as

k(x,x′)=σf2exp⁡(−∥x−x′∥22ℓ2).(4)k(x,x') = \sigma_f^2 \exp\left(-\frac{\lVert x-x'\rVert^2}{2\ell^2}\right). \tag*{(4)}

Preliminaries on Control Barrier Functions

Consider the nonlinear control-affine system

x˙=f(x)+g(x)u,x(0)=x0,(5)\dot{x} = f(x) + g(x)u,\qquad x(0)=x_0, \tag*{(5)}

where x(t)∈X⊆Rnx(t) \in\mathcal{X} \subseteq\mathbb{R}^n and u(t)∈U⊆Rmu(t) \in\mathcal{U} \subseteq\mathbb{R}^m are the state and control input, respectively. The functions f:Rn→Rnf : \mathbb{R}^n \to\mathbb{R}^n and g:Rn→Rn×mg : \mathbb{R}^n \to\mathbb{R}^{n\times m} are locally Lipschitz continuous and U⊂Rm\mathcal{U} \subset\mathbb{R}^m is nonempty, compact, and convex. Let h:R≥0×X→Rh : \mathbb{R}_{\ge0} \times\mathcal{X} \to\mathbb{R} be a continuously differentiable function and define the time-varying safe set C(t):={x∈X:h(t,x)≥0}\mathcal{C}(t) := \{x \in\mathcal{X} : h(t,x) \ge0\}. The family of sets C(t)\mathcal{C}(t) is forward invariant if any closed-loop trajectory initialized at x(t0)∈C(t0)x(t_0) \in\mathcal{C}(t_0) satisfies x(t)∈C(t)x(t) \in\mathcal{C}(t) for all t≥t0t \ge t_0. The function hh is a time-varying CBF [23, 24] if there exists a function α∈K\alpha\in\mathcal{K} such that

sup⁡u∈Uh˙(t,x,u)≥−α(h(t,x)),(6)\sup_{u\in\mathcal{U}} \dot{h}(t,x,u) \ge-\alpha(h(t,x)), \tag*{(6)}

for all t≥0t \ge0 and x∈C(t)x \in\mathcal{C}(t), where h˙(t,x,u)\dot{h}(t,x,u) is the total derivative of hh along the trajectories of system (5):

h˙(t,x,u)=∂th(t,x)+∇xh(t,x)T(f(x)+g(x)u).(7)\dot{h}(t,x,u) = \partial_t h(t,x) + \nabla_x h(t,x)^\mathsf{T}(f(x)+g(x)u). \tag*{(7)}

Condition (6) ensures the existence of an admissible input u∈Uu \in\mathcal{U} satisfying the CBF inequality, whose enforcement guarantees forward invariance of the safe set C(t)\mathcal{C}(t). The time-invariant case [2] is recovered by setting h(t,x)=h(x)h(t,x)=h(x), so that ∂th(t,x)=0\partial_t h(t,x)=0 and C(t)=C\mathcal{C}(t)=\mathcal{C} for all t≥0t \ge0.

Problem Formulation: Distance-Certified Safe Control

Consider the control-affine system (5), and let P:X→WP:\mathcal{X}\to\mathcal{W} be a C2C^2 map associating the state xx with a point p=P(x)p=P(x) in the workspace W⊆Rnp\mathcal{W}\subseteq\mathbb{R}^{n_p}. Let O(t)⊂W\mathcal{O}(t)\subset\mathcal{W} denote the closed set occupied by possibly moving obstacles, with nonempty compact boundary ∂O(t)\partial\mathcal{O}(t). We first characterize the evolution of the obstacles.

Assumption 1: There exist known constants vˉo,vˉ∂≥0\bar{v}_o,\bar{v}_{\partial}\ge0, such that, for every s∈[0,t]s\in[0,t],

dH(O(t),O(s))≤vˉo(t−s),(8)d_{\mathrm{H}}(\mathcal{O}(t),\mathcal{O}(s))\leq\bar{v}_o(t-s), \tag*{(8)}
dH(∂O(t),∂O(s))≤vˉ∂(t−s),(9)d_{\mathrm{H}}(\partial\mathcal{O}(t),\partial\mathcal{O}(s))\leq\bar{v}_{\partial}(t-s), \tag*{(9)}

where dHd_{\mathrm{H}} is the Hausdorff distance (cf. Notation). □

Given a prescribed margin d>0d>0, the physical safe set is

Sd(t):={x∈X∣dist⁡(P(x),∂O(t))≥d, P(x)∉int⁡(O(t))}.(10)\mathcal{S}_d(t):=\{x\in\mathcal{X}\mid\operatorname{dist}(P(x),\partial\mathcal{O}(t))\geq d,\ P(x)\notin\operatorname{int}(\mathcal{O}(t))\}. \tag*{(10)}

The true obstacle geometry O(t)\mathcal{O}(t) is unknown to the controller. Instead, at discrete time instants tk=kTt_k=kT, with k∈Nk\in\mathbb{N} and sampling period T>0T>0, a perception system (e.g., a camera) provides a finite collection of NkN_k distinct boundary observations:

XNk:={p1k,…,pNkk}⊂∂O(tk).(11)X_N^k:=\{p_1^k,\ldots,p_{N_k}^k\}\subset\partial\mathcal{O}(t_k). \tag*{(11)}

To account for the discrepancy between the observed point cloud and the complete obstacle boundary, we introduce the following assumption.

Assumption 2 (complete-boundary coverage): For every k∈Nk\in\mathbb{N}, the point cloud XNkX_N^k (cf. (10)) satisfies

dH(XNk,∂O(tk))≤δNk,(12)d_{\mathrm{H}}(X_N^k,\partial\mathcal{O}(t_k))\leq\delta_N^k, \tag*{(12)}

for some δNk>0\delta_N^k>0, where dHd_{\mathrm{H}} is the Hausdorff distance. □

Problem 1 (Distance-Certified Safe Control): Given system (5), a nominal control policy unomu_{\mathrm{nom}}, a prescribed minimum distance d>0d>0, and a sequence of point-cloud observations {XNk}k∈N≥0\{X_N^k\}_{k\in\mathbb{N}_{\geq0}}, design a continuously differentiable time-varying safety function HH and a safety controller usafeu_{\mathrm{safe}} such that the certified set

C(t):={x∈X∣H(t,x)≥0}(13)\mathcal{C}(t):=\{x\in\mathcal{X}\mid H(t,x)\geq0\} \tag*{(13)}

satisfies, for all t≥0t\geq0, the following properties:

H(t,x)≥0 ⟹ dist⁡(P(x),∂O(t))≥d,(14)H(t,x)\geq0\ \Longrightarrow\ \operatorname{dist}(P(x),\partial\mathcal{O}(t))\geq d, \tag*{(14)}
x(0)∈C(0) ⟹ x(t)∈C(t).(15)x(0)\in\mathcal{C}(0)\ \Longrightarrow\ x(t)\in\mathcal{C}(t). \tag*{(15)}

Distance-Certified GP-CBF Safe Controller: Strategy Description and Safety Guarantee

To address Problem 1, we propose a safety control strategy that (i) constructs distance-certified GP-based spatial barriers from successive point-cloud observations, (ii) smoothly aggregates them into a time-varying safety function, and (iii) enforces the resulting safety certificate through a CBF-based QP. The complete procedure is summarized in Algorithm 1. Formally, at each acquisition instant, let qik(p)=∥p−pik∥2q_i^k(p)=\lVert p-p_i^k\rVert^2 denote the squared distance from an observation pik∈XNkp_i^k\in X_N^k. We consider a GP with constant prior mean m0>0m_0>0, signal variance σf2>0\sigma_f^2>0 and a prescribed negative value yobs<0y_{\mathrm{obs}}<0 assigned to each observed boundary point. Define

A:=(m0−yobs)σf2σf2+σn2(16)A:=\frac{(m_0-y_{\mathrm{obs}})\sigma_f^2}{\sigma_f^2+\sigma_n^2} \tag*{(16)}

with A>m0A>m_0. The spatial weights (16) parametrized by βk>0\beta_k>0 and ν>12\nu>\frac{1}{2}, smoothly combine individual GP representations associated with the observed boundary points pikp_i^k into a single point-cloud-dependent GP representation.

Algorithm 1 Distance-Certified GP-based Safe Controller

  • Require: unomu_{\mathrm{nom}}, AA, m0m_0, σf2\sigma_f^2, ν\nu, β\beta, τM\tau_M, DMD_M, dd, vˉ∂\bar{v}_{\partial}, γ>0\gamma> 0

  • Acquire initial scan (XN0,δN0)(X_N^0,\delta_N^0) and set ϑ0(0)=1\vartheta_0(0)=1

  • Initialize H(0,x)H(0,x)

  • if a new scan (XNk,δNk)(X_N^k,\delta_N^k) is available then

    • Compute spatial weights (with wik(pjk)=1i=jw_i^k(p_j^k)=1_{i=j})

wik(p)=(qik(p))−νe−βk(qik(p))∑j=1Nk(qjk(p))−νe−βk(qjk(p))(17)w_i^k(p)=\frac{(q_i^k(p))^{-\nu}e^{-\beta_k(q_i^k(p))}}{\sum_{j=1}^{N_k}(q_j^k(p))^{-\nu}e^{-\beta_k(q_j^k(p))}} \tag*{(17)}
  • Set Distance-Certified GP length scale

ℓk=(DM+δNk)2+log⁡Nkβk2log⁡(A/m0)(18)\ell_k=\frac{\sqrt{(D_M+\delta_N^k)^2+\frac{\log N_k}{\beta_k}}}{\sqrt{2\log(A/m_0)}} \tag*{(18)}
  • Compute GP barrier

μXNk(p)=m0−A∑i=1Nkwik(p)e−qik(p)2ℓk2(19)\mu_{X_N^k}(p)=m_0-A\sum_{i=1}^{N_k}w_i^k(p)e^{-\frac{q_i^k(p)}{2\ell_k^2}} \tag*{(19)}
  • end if

  • Update active scans and temporal weights

I(t)={i:0≤t−ti≤τM}(20)\mathcal{I}(t)=\{i:0\leq t-t_i\leq\tau_M\} \tag*{(20)}
ϑi(t)=χτM(t−ti)∑j∈I(t)χτM(t−tj)(21)\vartheta_i(t)=\frac{\chi_{\tau_M}(t-t_i)}{\sum_{j\in\mathcal{I}(t)}\chi_{\tau_M}(t-t_j)} \tag*{(21)}
  • Compute time-varying barrier function

H(t,x)=1m0∑i∈I(t)ϑi(t)μXNi(P(x))(22)H(t,x)=\frac{1}{m_0}\sum_{i\in\mathcal{I}(t)}\vartheta_i(t)\mu_{X_N^i}(P(x)) \tag*{(22)}
  • Compute safe control input

usafe(t,x)=arg⁡min⁡u∈U 12∥u−unom(t,x)∥2u_{\mathrm{safe}}(t,x)=\underset{u\in\mathcal{U}}{\arg\min}\ \frac{1}{2}\lVert u-u_{\mathrm{nom}}(t,x)\rVert^2
s.t.∂tH+∇xHT(f(x)+g(x)u)≥−γH(23)\text{s.t.}\quad\partial_tH+\nabla_xH^\mathsf{T}(f(x)+g(x)u)\geq-\gamma H \tag*{(23)}
  • Apply usafe(t,x)u_{\mathrm{safe}}(t,x) to system (5)

To account for obstacle motion, each scan-based certificate is assigned a validity horizon τM=(M+1)T\tau_M=(M+1)T, with M∈NM\in\mathbb{N}, and an inflated safety distance DM=d+vˉ∂τMD_M=d+\bar{v}_{\partial}\tau_M, where vˉ∂\bar{v}_{\partial} is the boundary motion bound in (9) and dd is the prescribed minimum safety distance we wish to guarantee. The GP length scale (18) is calibrated using DMD_M, the point-cloud coverage error δNk\delta_N^k, and the spatial weighting approximation. To ensure smooth updates, the algorithm retains the active scans (20) within the validity horizon τM\tau_M and combines their associated GP barriers through the temporal weights (20). These are obtained by normalizing the compactly supported window $\chi_{\tau_M}(s)$, defined by specializing the general definition introduced in the Notation paragraph of Section I` to τ=τM\tau=\tau_M. Unlike the spatial weights, which combine observations within a single scan, the temporal weights regulate the contributions of different scans over time. Since χτM(s)\chi_{\tau_M}(s) and its first derivative vanish at the endpoints of its support, barriers are smoothly introduced and removed. The resulting normalized composite barrier function

B(t,p)=1m0∑i∈I(t)ϑi(t)μXNi(p)(24)B(t,p)=\frac{1}{m_0}\sum_{i\in\mathcal{I}(t)}\vartheta_i(t)\mu_{X_N^i}(p) \tag*{(24)}

is then mapped into the state space through P(x)P(x), yielding the time-varying barrier function H(t,x)H(t,x), defined in (22). At every time instant, H(t,x)H(t,x) is incorporated into the CBF-QP (23), which computes the safe control input by minimally modifying the nominal policy unom(t,x)u_{\mathrm{nom}}(t,x) while enforcing the time-varying barrier condition. The overall architecture is also illustrated in Figure 1.

Figure 1. Proposed GP-CBF safety controller structure.

Theorem 1 establishes the corresponding geometric certification and forward invariance guarantees. The theoretical analysis supporting these results is presented in § 5.

Theorem 1 (Distance-Certified GP-based Safe Controller): Consider system (5) and let Assumptions 1 and 2 hold for the obstacle model and acquisition system of § 3. Let H(t,x)H(t,x) be constructed by Algorithm 1 with d>0d>0, A>m0A>m_0,

βk>0\beta_k > 0, ν>12\nu> \frac{1}{2}, M≥1M \ge1 and γ>0\gamma> 0. Assume the CBF-QP (22) is feasible for all t≥0t \ge0 and generates unique, forward-complete closed-loop trajectories for all x(0)∈C(0)x(0) \in C(0). Then, the following hold.

  1. Geometric Certification: The barrier HH is continuously differentiable and satisfies

H(t,x)≥0⇒dist⁡(P(x),∂O(t))≥d,(25)H(t,x) \ge0 \Rightarrow\operatorname{dist}(P(x),\partial O(t)) \ge d, \tag*{(25)}

for all t≥0t \ge0 and x∈Xx \in\mathcal{X}.

  1. Forward Invariance: For every x(0)∈C(0)x(0) \in C(0), the closed-loop system under usafe(t,x)u_{\mathrm{safe}}(t,x) satisfies

H(t,x(t))≥e−γtH(0,x(0))≥0,(26)H(t,x(t)) \ge e^{-\gamma t}H(0,x(0)) \ge0, \tag*{(26)}

ensuring forward invariance of the safe set C(t)C(t). Specifically, if P(x(0))∉int⁡(O(0))P(x(0)) \notin\operatorname{int}(O(0)), then x(t)∈Sd(t)x(t) \in S_d(t) for all t≥0.t \ge0.

The proof of Theorem 1 is provided in Section V-D.

Theoretical Analysis

In this section, we establish the theoretical properties underlying the proposed approach. Section V-A proves that the barrier in (18) is the posterior mean of a valid GP induced by a suitable kernel function, while Section V-B derives the closed-form minimum-distance certificate. Section V-C extends the geometric guarantee to moving obstacles and characterizes the time-varying barrier obtained by the temporal combination of stored GP barriers. Finally, Section V-D provides the proof of Theorem 1, establishing forward invariance of the GP-based safe set, which guarantees a minimum distance from the true obstacle boundary.

GP Construction and Kernel Validity

To analyze the GP construction in Algorithm 1, we consider a fixed point cloud XNkX_N^k at a given acquisition instant tkt_k. In the following, we omit the scan index kk for notational simplicity, and we write XN:=XNkX_N := X_N^k, N:=NkN := N_k, β:=βk\beta:= \beta_k, δN:=δNk\delta_N := \delta_N^k, and ∂O:=∂O(tk)\partial O := \partial O(t_k). The time dependence will be restored in Section V-C when analyzing the temporal combination of GP barriers. Consider a single boundary observation pi∈XNp_i \in X_N with label yobs<0y_{\mathrm{obs}} < 0. For a GP with constant prior mean m0>0m_0 > 0, squared-exponential kernel as in (4), and length scale ℓ>0\ell> 0, the single-observation posterior mean obtained from (2) is

μi(p)=m0−Aexp⁡(−∥p−pi∥22ℓ2),(27)\mu_i(p) = m_0 - A \exp\left(-\frac{\lVert p-p_i\rVert^2}{2\ell^2}\right), \tag*{(27)}

where AA is defined in (15). Extending this construction to multiple observations is not straightforward. Indeed, the pointwise minimum of individual posterior means provides a representation associated with the nearest observed point, but it is generally nonsmooth where the nearest point changes. Instead, to obtain the differentiable posterior mean in (18), we introduce the smooth spatial weighting scheme in (16) which satisfies the cardinality condition wi(pj)=1i=jw_i(p_j)=1_{i=j}. The corresponding point-cloud-dependent covariance kernel is defined as

kXN(p,p′)=σf2e−∥p−p′∥22ℓ2∑i=1Nwi(p)wi(p′).(28)k_{X_N}(p,p')=\sigma_f^2 e^{-\frac{\lVert p-p'\rVert^2}{2\ell^2}}\sum_{i=1}^{N}w_i(p)w_i(p'). \tag*{(28)}

Such a kernel allows us to interpret the weighted combination of individual posteriors in (18) as a rigorous GP posterior mean, as formalized in the following result.

Proposition 1 (Validity of the GP Construction). Consider a fixed point cloud XNX_N and the kernel kXNk_{X_N} defined in (27). Then, kXNk_{X_N} is symmetric and positive semidefinite, and its observation Gram matrix satisfies KXN=σf2INK_{X_N}=\sigma_f^2 I_N. Moreover, conditioning the corresponding GP on (10) yields the posterior mean defined in (18). □

The proof of Proposition 1 is provided in Appendix A.

Remark 1: Thanks to the diagonal structure of the Gram matrix, the GP conditioning does not require the inversion of a dense covariance matrix, and (18) can be expressed as a weighted sum of the individual posteriors in (26). □

Remark 2: The geometric certification of the GP superlevel set must account for two sources of approximation: the smooth spatial weighting, which approximates the nearest-observation envelope, and the finite point-cloud sampling, which introduces a discrepancy with respect to the complete obstacle boundary. Both effects are accounted for in the calibration of the GP length scale ℓ\ell in (17), whose minimum-distance guarantee will be established in Lemma 2. □

Closed-Form Minimum-Distance Certificate

We now establish that the closed-form calibration of the GP length scale ℓ\ell in (17) guarantees the prescribed minimum distance from the true obstacle boundary. To account for the approximation introduced by the smooth weighting (16), the calibration incorporates the selection-error bound

ϵβ,N:=log⁡Nβ,(29)\epsilon_{\beta,N}:=\frac{\log N}{\beta}, \tag*{(29)}

which bounds the discrepancy between the weighted and nearest-observation squared distances. The following lemma establishes the corresponding bound on the weighted Gaussian terms in the posterior mean (18).

Lemma 1 (Gaussian-Envelope Bound). Consider the spatial weights wi(p)w_i(p) in (16) and let q⋆(p):=min⁡iqi(p)=dist⁡(p,XN)2q_\star(p):=\min_i q_i(p)=\operatorname{dist}(p,X_N)^2. Then, for every p∈Wp\in\mathcal{W} and ℓ>0\ell>0, it holds

e−q⋆(p)+ϵβ,N2ℓ2≤∑i=1Nwi(p)e−qi(p)2ℓ2≤e−q⋆(p)2ℓ2.(30)e^{-\frac{q_\star(p)+\epsilon_{\beta,N}}{2\ell^2}} \leq\sum_{i=1}^{N}w_i(p)e^{-\frac{q_i(p)}{2\ell^2}} \leq e^{-\frac{q_\star(p)}{2\ell^2}}. \tag*{(30)}

The proof of Lemma 1 is provided in Appendix B.

Figure 2 illustrates the effects of spatial weighting and finite boundary sampling on the GP zero-level contours. For a fixed point cloud, increasing β\beta reduces the selection error bound in (28), bringing the GP contour closer to the one-point minimum envelope. The negative deviations shown in the lower panels indicate that the resulting contours lie inside the prescribed minimum distance region. These effects motivate the length-scale calibration derived in the following result, which accounts for both spatial weighting and finite boundary coverage.

Lemma 2 (Minimum-Distance Certificate). Let Assumption 2 hold for a point cloud XNX_N. Consider the GP posterior μXN\mu_{X_N} in (18). Given any prescribed margin d>0d>0, choosing

ℓ≥ℓsafe:=(d+δN)2+ϵβ,N2log⁡(A/m0)(31)\ell\geq\ell_{\mathrm{safe}}:=\frac{\sqrt{(d+\delta_N)^2+\epsilon_{\beta,N}}}{\sqrt{2\log(A/m_0)}} \tag*{(31)}

guarantees that, for every p∈Wp\in\mathcal{W},

μXN(p)≥0⟹dist⁡(p,∂O)≥d.\mu_{X_N}(p)\geq0\quad\Longrightarrow\quad\operatorname{dist}(p,\partial\mathcal{O})\geq d.

The proof of Lemma 2 is provided in Appendix C.

Essentially, (30) in Lemma 2 provides a closed-form calibration of the GP length scale that guarantees a prescribed minimum distance dd from the complete obstacle boundary. We highlight that choosing d=DMd=D_M in (30) recovers the calibration of the GP length scale ℓ\ell in (17) adopted in Algorithm 1, extending the guarantee to the case of moving obstacles as will be established in the following subsection.

Figure 2. Effect of spatial weighting and boundary sampling on the (uncalibrated) GP zero-level contours for N=24N = 24 (left) and N=96N = 96 (right) observations. Top panels: (Uncalibrated) GP contours for two values of β\beta, compared with the ideal dd-offset envelope and the one-point minimum envelope. Bottom panels: Radial deviations from the ideal dd-offset for increasing values of β\beta.

Time-Varying Barrier Construction

We now extend the minimum-distance certificate derived in § 5-B (cf. Lemma 2) to the case of moving obstacles by relying on the temporal combination of stored GP barriers. We assume that each acquired point cloud is expressed in a common reference frame and kept fixed after acquisition. For each scan at time tkt_k, define the normalized GP barrier

bk(p)=μXNk(p)m0,(32)b_k(p) = \frac{\mu_{X_N^k}(p)}{m_0}, \tag*{(32)}

and its zero-superlevel set Kk:={p∈W∣bk(p)≥0}K_k := \{p \in\mathcal{W} \mid b_k(p) \ge0\}. Recall the validity horizon τM=(M+1)T\tau_M = (M + 1)T and the inflated safety distance DM=d+vˉ∂τMD_M = d + \bar{v}_{\partial}\tau_M introduced in the algorithm description. The following result establishes that the calibration (17) preserves the prescribed minimum distance throughout the validity horizon τM\tau_M.

Lemma 3 (Finite-Scan Boundary Validity): Let Assumptions 1 and 2 hold. Consider the GP barrier bkb_k constructed in (31) and its superlevel set KkK_k. Then,

dist⁡(p,∂O(t))≥DM−vˉ∂(t−tk)≥d(33)\operatorname{dist}(p, \partial\mathcal{O}(t)) \ge D_M - \bar{v}_{\partial}(t - t_k) \ge d \tag*{(33)}

for all p∈Kkp \in K_k and t∈[tk,tk+τM]t \in[t_k, t_k + \tau_M]. □

The proof of Lemma 3 is provided in Appendix D. Consequently, every active GP barrier retains its minimum-distance certificate despite obstacle motion. This property, together with the smooth temporal composition of stored GP barriers defined in (23), provides the basis for the geometric certification established in Theorem 1.

Proof of Theorem 1

We first establish the regularity of the composite barrier. By construction, the spatial weights (16), with ν>12\nu> \frac{1}{2}, and the corresponding GP posterior mean (18) are continuously differentiable. Moreover, since M≥1M \ge1 and scans are acquired periodically, the denominator in (20) is strictly positive for all t>0t > 0. As the temporal window χτM\chi_{\tau_M} (cf. Notation) and its derivative vanish at both endpoints of its support [ti,ti+τM][t_i,t_i+\tau_M], the temporal weights (20) are continuously differentiable across scan updates. At t=0t=0, the initialization ϑ0(0)=1\vartheta_0(0)=1 provides the continuous extension. Consequently, each stored barrier bib_i is continuously differentiable by construction, resulting in the composite barrier BB and the corresponding safety function HH being continuously differentiable. Since the temporal weights are nonnegative and sum to one, the composite barrier satisfies

B(t,p)=∑i∈I(t)ϑi(t)bi(p)≤max⁡i∈I(t)bi(p).(34)B(t,p)=\sum_{i\in\mathcal{I}(t)}\vartheta_i(t)b_i(p)\leq\max_{i\in\mathcal{I}(t)}b_i(p). \tag*{(34)}

Thus, B(t,p)≥0B(t,p)\geq0 implies bi(p)≥0b_i(p)\geq0 for at least one active scan i∈I(t)i\in\mathcal{I}(t). Since every active scan lies within its validity horizon, we can apply Lemma 3 to guarantee

H(t,x)≥0⇒dist⁡(P(x),∂O(t))≥d.(35)H(t,x)\geq0\Rightarrow\operatorname{dist}(P(x),\partial\mathcal{O}(t))\geq d. \tag*{(35)}

The same implication holds at t=0t=0 by the initialization of the composite barrier with the first scan, establishing (24) and, thus, concluding the proof of the geometric certification of the composite barrier.

Now, consider a closed-loop trajectory initialized in C(0)\mathcal{C}(0). By the assumed feasibility and well-posedness of the CBF-QP, the trajectory exists uniquely for all t≥0t\geq0. Applying the chain rule and the CBF condition in (22) yields

ddt(eγtH)=eγt(∂tH+∇xH⊤(f(x)+g(x)u)+γH)≥0.(36)\frac{d}{dt}\left(e^{\gamma t}H\right)=e^{\gamma t}\left(\partial_tH+\nabla_xH^\top(f(x)+g(x)u)+\gamma H\right)\geq0. \tag*{(36)}

Integrating the above inequality gives

H(t,x(t))≥e−γtH(0,x(0))≥0,(37)H(t,x(t))\geq e^{-\gamma t}H(0,x(0))\geq0, \tag*{(37)}

which establishes (25) and forward invariance of C(t)\mathcal{C}(t).

Finally, let p(t)=P(x(t))p(t)=P(x(t)), with p(0)∉int⁡(O(0))p(0)\notin\operatorname{int}(\mathcal{O}(0)). Exploiting the obstacle motion bound (8a) (cf. Assumption 1) and the triangle inequality, we have, for every t≥st\geq s

dist⁡(p(t),O(t))≥dist⁡(p(s),O(s))−∥p(t)−p(s)∥−vˉo(t−s).(38)\operatorname{dist}(p(t),\mathcal{O}(t))\geq\operatorname{dist}(p(s),\mathcal{O}(s))-\lVert p(t)-p(s)\rVert-\bar{v}_o(t-s). \tag*{(38)}

By the initial conditions and the condition in (34), p(0)∉O(0)p(0)\notin\mathcal{O}(0). Suppose, by contradiction, that the trajectory p(t)p(t) intersects the obstacle, and let t⋆>0t_\star>0 be the first contact time. For every s∈[0,t⋆)s\in[0,t_\star), we have

dist⁡(p(s),O(s))=dist⁡(p(s),∂O(s))≥d>0.(39)\operatorname{dist}(p(s),\mathcal{O}(s))=\operatorname{dist}(p(s),\partial\mathcal{O}(s))\geq d>0. \tag*{(39)}

Letting s→t⋆s\to t_\star, and using continuity of p(t)p(t), we obtain dist⁡(p(t⋆),O(t⋆))≥d>0\operatorname{dist}(p(t_\star),\mathcal{O}(t_\star))\geq d>0. Applying (37) again from s=t⋆s=t_\star, and using continuity of p(t)p(t) we see that the trajectory cannot enter the obstacle for a sufficiently small time interval after t⋆t_\star, contradicting the assumption that p(t)p(t) intersects the obstacle at t⋆t_\star. Thus, p(t)∉O(t)p(t)\notin\mathcal{O}(t) for all t≥0t\geq0. Together with (24) this establishes x(t)∈Sd(t)x(t)\in\mathcal{S}_d(t) for all t≥0.□t\geq0. \square

Numerical Simulations

We evaluate the proposed approach on a disk-shaped unicycle robot navigating an unknown environment with static and moving obstacles of different geometries.

The robot state is x=(px,py,θ)⊤x = (p_x,p_y,\theta)^{\top}, where (px,py)⊤(p_x,p_y)^{\top} denotes the robot center position and θ\theta its orientation. The robot dynamics are given by p˙x=vcos⁡θ\dot{p}_x = v\cos\theta, p˙y=vsin⁡θ\dot{p}_y = v\sin\theta, θ˙=ω\dot{\theta} = \omega, with ∣v∣≤1|v| \leq1 m/s and ∣ω∣≤6|\omega| \leq6 rad/s. The robot has radius rb=0.15r_b = 0.15 m and protected point P(x)=(px,py)⊤+L(cos⁡θ,sin⁡θ)⊤P(x) = (p_x,p_y)^{\top} + L(\cos\theta,\sin\theta)^{\top}, with L=0.15L = 0.15 m. A minimum required distance of db=0.15d_b = 0.15 m between the robot body and the obstacle boundary is enforced by prescribing the protected-point distance d=db+rb+L=0.45d = d_b + r_b + L = 0.45 m. Starting from x(0)=(0,0,0)x(0) = (0,0,0), the robot must steer its lookahead point toward the target point (48,0)(48,0) m. The GP parameters are m0=1m_0 = 1, σf=1\sigma_f = 1, ν=2\nu= 2, β=395\beta= 395 m−2^{-2}. Scans are acquired every T=0.12T = 0.12 s, with M=6M = 6 and τM=(M+1)T=0.84\tau_M = (M + 1)T = 0.84 s. To illustrate Assumption 2, we simulate an overhead camera with complete boundary visibility. The workspace is discretized into square cells of side lc=0.20l_c = 0.20 m, with each boundary-intersecting cell providing an observation. This yields a constant coverage bound δNk=lc2≃0.28\delta_N^k = l_c\sqrt{2} \simeq0.28 m. The boundary and obstacle motion bounds are vˉ∂=vˉo=0.19\bar{v}_{\partial} = \bar{v}_o = 0.19 m/s, yielding (DM+δN)2+ϵβ,N=0.9018\sqrt{(D_M + \delta_N)^2 + \epsilon_{\beta,N}} = 0.9018 m. Since boundary observations are assigned the fixed label yobs=−1y_{\mathrm{obs}} = -1, we set the noise variance to σn2=0\sigma_n^2 = 0. The nominal controller assigns the protected-point velocity using a workspace proportional controller with kg=0.8k_g = 0.8 and maximum velocity vg=0.4v_g = 0.4 m/s. The corresponding physical inputs are obtained through the protected-point Jacobian and saturated at the actuator limits. The safety controller solves (22) in the scaled input u~=(v,Lω)\tilde{u} = (v,L\omega). We simulate the proposed safety controller for γ∈{0.1,0.5,10}\gamma\in\{0.1, 0.5, 10\}. Trajectories are integrated using the RK45 solver of SciPy, with relative and absolute tolerances 10−710^{-7} and 10−910^{-9}, respectively, and a maximum step of 0.02 s. Figure 3 shows that the proposed strategy maintains nonnegative barrier values and preserves the prescribed minimum distance from the obstacles for all considered values of γ\gamma. Figure 4 illustrates the evolution of the GP barrier in the presence of moving obstacles, highlighting the adaptation of the certified safe region to the changing environment.

Figure 3. Left: Time evolution of the GP-CBF H(t,x(t))H(t,x(t)) Right: Protected-point distance for γ=0.1\gamma=0.1 (blue), γ=0.5\gamma=0.5 (orange), and γ=10\gamma=10 (green).

Map of the environment and snapshots of the dynamic environment with corresponding GP barrier

Figure 4. Top panel: Map of the environment. Bottom panels: Snapshots of the dynamic environment and corresponding GP barrier for γ=0.5\gamma= 0.5. Black dots denote boundary observations, black curves represent the GP zero-level sets, and blue lines show the last 12 seconds of the robot trajectory.

Conclusions

We proposed a distance-certified safety control framework for unknown and dynamic environments based on GP barriers constructed from point-cloud observations. We introduced a novel covariance kernel that enables efficient computation of GP-based safety certificates. By accounting for spatial approximation errors and finite point-cloud coverage, we derived a closed-form calibration of the GP length scale that guarantees a prescribed minimum distance from the time-varying obstacle boundaries. The resulting safety controller guarantees forward invariance of the certified safe set and preservation of the prescribed distance.

Proof of Proposition 1

Consider NN independent zero-mean Gaussian processes Gi∼GP(0,kSE)G_i \sim\mathcal{GP}(0,k_{\mathrm{SE}}), i=1,…,Ni=1,\ldots,N, with kSEk_{\mathrm{SE}} as in (4). Construct the random field

F(p):=m0+∑i=1Nwi(p)Gi(p).(40)F(p) := m_0 + \sum_{i=1}^{N} w_i(p)G_i(p). \tag*{(40)}

Since the weights are deterministic and the GiG_i are independent, FF is a GP with mean m0m_0 and covariance Cov⁡(F(p),F(p′))=kSE(p,p′)∑i=1Nwi(p)wi(p′)=kXN(p,p′)\operatorname{Cov}(F(p),F(p')) = k_{\mathrm{SE}}(p,p')\sum_{i=1}^{N}w_i(p)w_i(p') = k_{XN}(p,p'). Thus, kXNk_{XN} is symmetric and positive-semidefinite. Moreover, exploiting the cardinality property of the spatial weights (17), we obtain

kXN(pj,pk)=σf2∑i=1Nwi(pj)wi(pk)=σf21j=k,(41)k_{XN}(p_j,p_k)=\sigma_f^2\sum_{i=1}^{N}w_i(p_j)w_i(p_k)=\sigma_f^2\mathbf{1}_{j=k}, \tag*{(41)}
kXN(p,pj)=σf2wj(p)e−qj(p)2ℓ2.(42)k_{XN}(p,p_j)=\sigma_f^2w_j(p)e^{-\frac{q_j(p)}{2\ell^2}}. \tag*{(42)}

The equality (41) implies KXN=σf2INK_{XN}=\sigma_f^2I_N. Substituting both (41) and (42) into the GP conditioning formula (cf. (2)), with X=[p1,…,pN]TX=[p_1,\ldots,p_N]^{\mathsf{T}} and all outputs equal to yobs<0y_{\mathrm{obs}}<0, yields

μXN(p)=m0+yobs−m0σf2+σn2∑i=1NkXN(p,pi)=m0−A∑i=1Nwi(p)e−qi(p)2ℓ2,(43)\mu_{XN}(p)=m_0+\frac{y_{\mathrm{obs}}-m_0}{\sigma_f^2+\sigma_n^2}\sum_{i=1}^{N}k_{XN}(p,p_i)=m_0-A\sum_{i=1}^{N}w_i(p)e^{-\frac{q_i(p)}{2\ell^2}}, \tag*{(43)}

which recovers (19) and completes the proof. □

Proof of Lemma 1

At observation points pi∈XNp_i\in X_N, the claim follows directly from the cardinality property wi(pj)=1i=jw_i(p_j)=\mathbf{1}_{i=j}. Otherwise, fix p∈W∖XNp\in W\setminus X_N and define the following auxiliary quantities

Δi:=qi−q⋆,ri:=(q⋆qi)ν,Z:=∑i=1Nrie−βΔi.(44)\Delta_i:=q_i-q_\star,\qquad r_i:=\left(\frac{q_\star}{q_i}\right)^\nu,\qquad Z:=\sum_{i=1}^{N}r_i e^{-\beta\Delta_i}. \tag*{(44)}

Since qi≥q⋆>0q_i\ge q_\star>0 for all ii, we have ri∈(0,1]r_i\in(0,1]. Moreover, at least one i⋆i_\star satisfies qi⋆=q⋆q_{i_\star}=q_\star, implying ri⋆=1r_{i_\star}=1 and Δi⋆=0\Delta_{i_\star}=0; hence Z≥1Z\ge1. The weights wi(p)w_i(p) in (17) can be thus rewritten as wi(p)=rie−βΔiZw_i(p)=\frac{r_i e^{-\beta\Delta_i}}{Z}. Taking logarithms and introducing the Shannon entropy H(w):=−∑i=1Nwilog⁡wiH(w):=-\sum_{i=1}^{N}w_i\log w_i, we obtain

β∑i=1NwiΔi=H(w)+∑i=1Nwilog⁡ri−log⁡Z≤(a)log⁡N.(45)\beta\sum_{i=1}^{N}w_i\Delta_i=H(w)+\sum_{i=1}^{N}w_i\log r_i-\log Z\overset{(a)}{\le}\log N. \tag*{(45)}

where in (a) we use ri≤1r_i\le1 and Z≥1Z\ge1, and the known upper bound on the entropy of a discrete probability distribution over NN elements. Also, since z↦e−z/(2ℓ2)z\mapsto e^{-z/(2\ell^2)} is convex and decreasing,

e−q⋆+εβ,N2ℓ2≤(a)e−∑iwiqi2ℓ2≤(b)∑i=1Nwie−qi2ℓ2≤(c)e−q⋆2ℓ2.(46)e^{-\frac{q_\star+\varepsilon_{\beta,N}}{2\ell^2}}\overset{(a)}{\le}e^{-\frac{\sum_i w_iq_i}{2\ell^2}}\overset{(b)}{\le}\sum_{i=1}^{N}w_i e^{-\frac{q_i}{2\ell^2}}\overset{(c)}{\le}e^{-\frac{q_\star}{2\ell^2}}. \tag*{(46)}

where (a) follows from (45) and the monotonicity of the exponential, in (b) we use Jensen’s inequality, and (c) follows from qi≥q⋆q_i\ge q_\star and ∑iwi(p)=1\sum_iw_i(p)=1. □

Proof of Lemma 2

Let p∈Wp\in W satisfy μXN(p)≥0\mu_{XN}(p)\ge0. Recall that q⋆(p)=dist⁡(p,XN)2q_\star(p)=\operatorname{dist}(p,X_N)^2. Combining the Gaussian-envelope bound (30) with the definition of the GP posterior mean (19), we get

e−q⋆(p)+εβ,N2ℓ2≤∑i=1Nwi(p)e−qi(p)2ℓ2≤m0A.(47)e^{-\frac{q_\star(p)+\varepsilon_{\beta,N}}{2\ell^2}}\le\sum_{i=1}^{N}w_i(p)e^{-\frac{q_i(p)}{2\ell^2}}\le\frac{m_0}{A}. \tag*{(47)}

Taking logarithms and using the condition (31) yields

q⋆(p)≥2ℓ2log⁡(A/m0)−εβ,N≥(d+δN)2.(48)q_\star(p)\ge2\ell^2\log(A/m_0)-\varepsilon_{\beta,N}\ge(d+\delta_N)^2. \tag*{(48)}

Therefore, by Assumption 2,

dist⁡(p,∂O)≥q⋆−δN≥d+δN−δN=d,(49)\operatorname{dist}(p,\partial O)\ge\sqrt{q_\star}-\delta_N\ge d+\delta_N-\delta_N=d, \tag*{(49)}

which concludes the proof. □

Proof of Lemma 3

Fix k≥0k \ge0 and let p∈Kkp \in K_k. By definition (31), bk(p)≥0b_k(p) \ge0 implies μXNk(p)≥0\mu_{X_N^k}(p) \ge0. Since the calibration (17) corresponds to (30) with d=DMd = D_M, Lemma 2 implies

dist⁡(p,∂O(tk))≥DM.(50)\operatorname{dist}(p, \partial O(t_k)) \ge D_M. \tag*{(50)}

Moreover, for any t∈[tk,tk+τM]t \in[t_k, t_k + \tau_M]

dist⁡(p,∂O(t))≥(a)dist⁡(p,∂O(tk))−vˉ∂(t−tk)≥(b)DM−vˉ∂(t−tk)≥(c)d,(51)\operatorname{dist}(p, \partial O(t)) \overset{(a)}{\ge} \operatorname{dist}(p, \partial O(t_k)) - \bar{v}_{\partial}(t - t_k) \overset{(b)}{\ge} D_M - \bar{v}_{\partial}(t - t_k) \overset{(c)}{\ge} d, \tag*{(51)}

where (a)(a) follows from Assumption 1, (b)(b) follows from (49), (c)(c) follows from both the definition of DMD_M in the algorithm description, and the inequality t−tk≤τMt - t_k \le\tau_M. Since p∈Kkp \in K_k was arbitrary, the bound (50) holds for all p∈Kkp \in K_k. □\square

References

  1. [1]K.-C. Hsu, H. Hu, and J. F. Fisac, “The safety filter: A unified view of safety-critical control in autonomous systems,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 7, 2023.
  2. [2]A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,” IEEE transactions on automatic control, vol. 62, no. 8, pp. 3861–3876, 2016.DOI
  3. [3]A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control barrier functions: Theory and applications,” in 2019 18th European control conference (ECC), IEEE, 2019, pp. 3420–3431.arxiv.org/abs/1903.11199
  4. [4]L. Wang, A. D. Ames, and M. Egerstedt, “Safety barrier certificates for collisions-free multirobot systems,” IEEE Transactions on Robotics, vol. 33, no. 3, pp. 661–674, 2017.DOI
  5. [5]F. Ferraguti, C. T. Landi, A. Singletary, H.-C. Lin, A. Ames, C. Secchi, and M. Bonfè, “Safety and efficiency in robotics: The control barrier functions approach,” IEEE Robotics & Automation Magazine, vol. 29, no. 3, pp. 139–151, 2022.DOI
  6. [6]A. Singletary, S. Kolathaya, and A. D. Ames, “Safety-critical kinematic control of robotic systems,” IEEE Control Systems Letters, vol. 6, pp. 139–144, 2021.
  7. [7]C. T. Landi, F. Ferraguti, S. Costi, M. Bonfè, and C. Secchi, “Safety barrier functions for human-robot interaction with industrial manipulators,” in 2019 18th European Control Conference (ECC). IEEE, 2019, pp. 2565–2570.DOI
  8. [8]A. Alan, A. J. Taylor, C. R. He, A. D. Ames, and G. Orosz, “Control barrier functions and input-to-state safety with application to automated vehicles,” IEEE Transactions on Control Systems Technology, vol. 31, no. 6, pp. 2744–2759, 2023.
  9. [9]A. Robey, H. Hu, L. Lindemann, H. Zhang, D. V. Dimarogonas, S. Tu, and N. Matni, “Learning control barrier functions from expert demonstrations,” in 2020 59th IEEE Conference on Decision and Control (CDC), 2020, pp. 3717–3724.
  10. [10]S. Dean, A. Taylor, R. Cosner, B. Recht, and A. Ames, “Guaranteeing safety of learned perception modules via measurement-robust control barrier functions,” in Conference on Robot Learning. PMLR, 2021, pp. 654–670.
  11. [11]K. Long, C. Qian, J. Cortés, and N. Atanasov, “Learning barrier functions with memory for robust safe navigation,” IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4931–4938, 2021.
  12. [12]M. De Sa, P. Kotaru, and K. Sreenath, “Point cloud-based control barrier function regression for safe and efficient vision-based control,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 366–372.
  13. [13]B. Dai, R. Khorrambakht, P. Krishnamurthy, and F. Khorrami, “Navigating through point clouds: Safe navigation using point cloud based control barrier functions,” IEEE Robotics and Automation Letters, vol. 9, no. 9, pp. 7731–7738, 2024.
  14. [14]P. Lutkus, M. S. Chong, and L. Lindemann, “Nonconvex obstacle avoidance using efficient sampling-based distance functions,” in 2025 IEEE 64th Conference on Decision and Control (CDC). IEEE, 2025, pp. 6784–6791.
  15. [15]C. K. Williams and C. E. Rasmussen, Gaussian processes for machine learning. MIT press, Cambridge, MA, 2006, vol. 2, no. 3.DOI
  16. [16]Y. Hu, J. Fu, and G. Wen, “Decentralized robust collision-avoidance for cooperative multirobot systems: A gaussian process-based control barrier function approach,” IEEE Transactions on Control of Network Systems, vol. 10, no. 2, pp. 706–717, 2022.
  17. [17]J. Li, Q. Liu, W. Jin, J. Qin, and S. Hirche, “Robust safe learning and control in an unknown environment: An uncertainty-separated control barrier function approach,” IEEE Robotics and Automation Letters, vol. 8, no. 10, pp. 6539–6546, 2023.
  18. [18]S. Keyumarsi, M. W. S. Atman, and A. Gusrialdi, “Lidar-based online control barrier function synthesis for safe navigation in unknown environments,” IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1043–1050, 2023.
  19. [19]Y. Zhang, L. Wen, X. Yao, Z. Bing, L. Kong, W. He, and A. Knoll, “Real-time adaptive safety-critical control with gaussian processes in high-order uncertain models,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 14763–14769.
  20. [20]X. Yin, C. Liang, Y. Guo, and J. Mei, “Dynamic log-gaussian process control barrier function for safe robotic navigation in dynamic environments,” in 2025 IEEE 64th Conference on Decision and Control (CDC). IEEE, 2025, pp. 3115–3120.
  21. [21]G. Bahati, R. M. Bena, and A. D. Ames, “Dynamic safety in complex environments: Synthesizing safety filters with poisson’s equation,” arXiv preprint arXiv:2505.06794, 2025.
  22. [22]G. Gu and K. Sreenath, “Safety-critical control of a planar quadrotor,” in 2016 American control conference (ACC). IEEE, 2016, pp. 2252–2258.
  23. [23]L. Lindemann and D. V. Dimarogonas, “Control barrier functions for signal temporal logic tasks,” IEEE control systems letters, vol. 3, no. 1, pp. 96–101, 2021.
  24. [24]X. Xu, “Constrained control of input–output linearizable systems using control sharing barrier functions,” Automatica, vol. 87, pp. 195–201, 2018.
  25. [25]M. Igarashi, I. Tezuka, and H. Nakamura, “Time-varying control barrier function and its application to environment-adaptive human assist control,” IFAC-PapersOnLine, vol. 52, no. 16, pp. 735–740, 2019.
  26. [26]J. Huang, Z. Liu, J. Zeng, X. Chi, and H. Su, “Obstacle avoidance for unicycle-modelled mobile robots with time-varying control barrier functions,” in 49th annual Conference of the IEEE Industrial Electronics Society (IECON), 2023, pp. 1–6.
  27. [27]B. Dai, R. Khorrambakht, P. Krishnamurthy, and F. Khorrami, “Differentiable optimization based time-varying control barrier functions for dynamic obstacle avoidance,” Robotics and Autonomous Systems, p. 105182, 2025.
  28. [28]T. G. Molnar and A. D. Ames, “Composing control barrier functions for complex safety specifications,” IEEE Control Systems Letters, vol. 7, pp. 3615–3620, 2023.
  29. [29]A. Safari and J. B. Hoagg, “Time-varying soft-maximum barrier functions for safety in unmapped and dynamic environments,” IEEE Transactions on Control Systems Technology, 2025.

Paper details

vlnc:W_hrs3uuut
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents